| 1 | /* origin: FreeBSD /usr/src/lib/msun/src/e_rem_pio2.c */ |
| 2 | /* |
| 3 | * ==================================================== |
| 4 | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
| 5 | * |
| 6 | * Developed at SunSoft, a Sun Microsystems, Inc. business. |
| 7 | * Permission to use, copy, modify, and distribute this |
| 8 | * software is freely granted, provided that this notice |
| 9 | * is preserved. |
| 10 | * ==================================================== |
| 11 | * |
| 12 | * Optimized by Bruce D. Evans. |
| 13 | */ |
| 14 | /* __rem_pio2(x,y) |
| 15 | * |
| 16 | * return the remainder of x rem pi/2 in y[0]+y[1] |
| 17 | * use __rem_pio2_large() for large x |
| 18 | */ |
| 19 | |
| 20 | #include "libm.h" |
| 21 | |
| 22 | #if FLT_EVAL_METHOD==0 || FLT_EVAL_METHOD==1 |
| 23 | #define EPS DBL_EPSILON |
| 24 | #elif FLT_EVAL_METHOD==2 |
| 25 | #define EPS LDBL_EPSILON |
| 26 | #endif |
| 27 | |
| 28 | /* |
| 29 | * invpio2: 53 bits of 2/pi |
| 30 | * pio2_1: first 33 bit of pi/2 |
| 31 | * pio2_1t: pi/2 - pio2_1 |
| 32 | * pio2_2: second 33 bit of pi/2 |
| 33 | * pio2_2t: pi/2 - (pio2_1+pio2_2) |
| 34 | * pio2_3: third 33 bit of pi/2 |
| 35 | * pio2_3t: pi/2 - (pio2_1+pio2_2+pio2_3) |
| 36 | */ |
| 37 | static const double |
| 38 | toint = 1.5/EPS, |
| 39 | pio4 = 0x1.921fb54442d18p-1, |
| 40 | invpio2 = 6.36619772367581382433e-01, /* 0x3FE45F30, 0x6DC9C883 */ |
| 41 | pio2_1 = 1.57079632673412561417e+00, /* 0x3FF921FB, 0x54400000 */ |
| 42 | pio2_1t = 6.07710050650619224932e-11, /* 0x3DD0B461, 0x1A626331 */ |
| 43 | pio2_2 = 6.07710050630396597660e-11, /* 0x3DD0B461, 0x1A600000 */ |
| 44 | pio2_2t = 2.02226624879595063154e-21, /* 0x3BA3198A, 0x2E037073 */ |
| 45 | pio2_3 = 2.02226624871116645580e-21, /* 0x3BA3198A, 0x2E000000 */ |
| 46 | pio2_3t = 8.47842766036889956997e-32; /* 0x397B839A, 0x252049C1 */ |
| 47 | |
| 48 | /* caller must handle the case when reduction is not needed: |x| ~<= pi/4 */ |
| 49 | int __rem_pio2(double x, double *y) |
| 50 | { |
| 51 | 	union {double f; uint64_t i;} u = {x}; |
| 52 | 	double_t z,w,t,r,fn; |
| 53 | 	double tx[3],ty[2]; |
| 54 | 	uint32_t ix; |
| 55 | 	int sign, n, ex, ey, i; |
| 56 | |
| 57 | 	sign = u.i>>63; |
| 58 | 	ix = u.i>>32 & 0x7fffffff; |
| 59 | 	if (ix <= 0x400f6a7a) { /* |x| ~<= 5pi/4 */ |
| 60 | 		if ((ix & 0xfffff) == 0x921fb) /* |x| ~= pi/2 or 2pi/2 */ |
| 61 | 			goto medium; /* cancellation -- use medium case */ |
| 62 | 		if (ix <= 0x4002d97c) { /* |x| ~<= 3pi/4 */ |
| 63 | 			if (!sign) { |
| 64 | 				z = x - pio2_1; /* one round good to 85 bits */ |
| 65 | 				y[0] = z - pio2_1t; |
| 66 | 				y[1] = (z-y[0]) - pio2_1t; |
| 67 | 				return 1; |
| 68 | 			} else { |
| 69 | 				z = x + pio2_1; |
| 70 | 				y[0] = z + pio2_1t; |
| 71 | 				y[1] = (z-y[0]) + pio2_1t; |
| 72 | 				return -1; |
| 73 | 			} |
| 74 | 		} else { |
| 75 | 			if (!sign) { |
| 76 | 				z = x - 2*pio2_1; |
| 77 | 				y[0] = z - 2*pio2_1t; |
| 78 | 				y[1] = (z-y[0]) - 2*pio2_1t; |
| 79 | 				return 2; |
| 80 | 			} else { |
| 81 | 				z = x + 2*pio2_1; |
| 82 | 				y[0] = z + 2*pio2_1t; |
| 83 | 				y[1] = (z-y[0]) + 2*pio2_1t; |
| 84 | 				return -2; |
| 85 | 			} |
| 86 | 		} |
| 87 | 	} |
| 88 | 	if (ix <= 0x401c463b) { /* |x| ~<= 9pi/4 */ |
| 89 | 		if (ix <= 0x4015fdbc) { /* |x| ~<= 7pi/4 */ |
| 90 | 			if (ix == 0x4012d97c) /* |x| ~= 3pi/2 */ |
| 91 | 				goto medium; |
| 92 | 			if (!sign) { |
| 93 | 				z = x - 3*pio2_1; |
| 94 | 				y[0] = z - 3*pio2_1t; |
| 95 | 				y[1] = (z-y[0]) - 3*pio2_1t; |
| 96 | 				return 3; |
| 97 | 			} else { |
| 98 | 				z = x + 3*pio2_1; |
| 99 | 				y[0] = z + 3*pio2_1t; |
| 100 | 				y[1] = (z-y[0]) + 3*pio2_1t; |
| 101 | 				return -3; |
| 102 | 			} |
| 103 | 		} else { |
| 104 | 			if (ix == 0x401921fb) /* |x| ~= 4pi/2 */ |
| 105 | 				goto medium; |
| 106 | 			if (!sign) { |
| 107 | 				z = x - 4*pio2_1; |
| 108 | 				y[0] = z - 4*pio2_1t; |
| 109 | 				y[1] = (z-y[0]) - 4*pio2_1t; |
| 110 | 				return 4; |
| 111 | 			} else { |
| 112 | 				z = x + 4*pio2_1; |
| 113 | 				y[0] = z + 4*pio2_1t; |
| 114 | 				y[1] = (z-y[0]) + 4*pio2_1t; |
| 115 | 				return -4; |
| 116 | 			} |
| 117 | 		} |
| 118 | 	} |
| 119 | 	if (ix < 0x413921fb) { /* |x| ~< 2^20*(pi/2), medium size */ |
| 120 | medium: |
| 121 | 		/* rint(x/(pi/2)) */ |
| 122 | 		fn = (double_t)x*invpio2 + toint - toint; |
| 123 | 		n = (int32_t)fn; |
| 124 | 		r = x - fn*pio2_1; |
| 125 | 		w = fn*pio2_1t; /* 1st round, good to 85 bits */ |
| 126 | 		/* Matters with directed rounding. */ |
| 127 | 		if (predict_false(r - w < -pio4)) { |
| 128 | 			n--; |
| 129 | 			fn--; |
| 130 | 			r = x - fn*pio2_1; |
| 131 | 			w = fn*pio2_1t; |
| 132 | 		} else if (predict_false(r - w > pio4)) { |
| 133 | 			n++; |
| 134 | 			fn++; |
| 135 | 			r = x - fn*pio2_1; |
| 136 | 			w = fn*pio2_1t; |
| 137 | 		} |
| 138 | 		y[0] = r - w; |
| 139 | 		u.f = y[0]; |
| 140 | 		ey = u.i>>52 & 0x7ff; |
| 141 | 		ex = ix>>20; |
| 142 | 		if (ex - ey > 16) { /* 2nd round, good to 118 bits */ |
| 143 | 			t = r; |
| 144 | 			w = fn*pio2_2; |
| 145 | 			r = t - w; |
| 146 | 			w = fn*pio2_2t - ((t-r)-w); |
| 147 | 			y[0] = r - w; |
| 148 | 			u.f = y[0]; |
| 149 | 			ey = u.i>>52 & 0x7ff; |
| 150 | 			if (ex - ey > 49) { /* 3rd round, good to 151 bits, covers all cases */ |
| 151 | 				t = r; |
| 152 | 				w = fn*pio2_3; |
| 153 | 				r = t - w; |
| 154 | 				w = fn*pio2_3t - ((t-r)-w); |
| 155 | 				y[0] = r - w; |
| 156 | 			} |
| 157 | 		} |
| 158 | 		y[1] = (r - y[0]) - w; |
| 159 | 		return n; |
| 160 | 	} |
| 161 | 	/* |
| 162 | 	 * all other (large) arguments |
| 163 | 	 */ |
| 164 | 	if (ix >= 0x7ff00000) { /* x is inf or NaN */ |
| 165 | 		y[0] = y[1] = x - x; |
| 166 | 		return 0; |
| 167 | 	} |
| 168 | 	/* set z = scalbn(|x|,-ilogb(x)+23) */ |
| 169 | 	u.f = x; |
| 170 | 	u.i &= (uint64_t)-1>>12; |
| 171 | 	u.i |= (uint64_t)(0x3ff + 23)<<52; |
| 172 | 	z = u.f; |
| 173 | 	for (i=0; i < 2; i++) { |
| 174 | 		tx[i] = (double)(int32_t)z; |
| 175 | 		z = (z-tx[i])*0x1p24; |
| 176 | 	} |
| 177 | 	tx[i] = z; |
| 178 | 	/* skip zero terms, first term is non-zero */ |
| 179 | 	while (tx[i] == 0.0) |
| 180 | 		i--; |
| 181 | 	n = __rem_pio2_large(tx,ty,(int)(ix>>20)-(0x3ff+23),i+1,1); |
| 182 | 	if (sign) { |
| 183 | 		y[0] = -ty[0]; |
| 184 | 		y[1] = -ty[1]; |
| 185 | 		return -n; |
| 186 | 	} |
| 187 | 	y[0] = ty[0]; |
| 188 | 	y[1] = ty[1]; |
| 189 | 	return n; |
| 190 | } |